Access to Elite Education, Wage Premium, and Social Mobility:

The Truth and Illusion of China’s College Entrance Exam∗

Ruixue Jia†and Hongbin Li‡

March 15, 2017

Abstract

This paper studies the returns to elite education and the implications of elite edu-

cation on elite formation and social mobility, exploiting an open elite education recruit-

ment system – China’s College Entrance Exam. We conduct annual national surveys of

around 40,000 college graduates during 2010-2015 to collect their scores at the college

entrance exam, job outcomes, and other individual and family characteristics. Ex-

ploiting a discontinuity in elite university eligibility around the cutoff scores, we find

a sizable wage premium of elite education but elite education eligibility does not nec-

essarily promise one’s entry into the elite class (measured by occupation, industry and

other non-wage benefits). While elite education eligibility does significantly affect one’s

mobility, it does not alter the influence of one’s parental background. We also find that

the wage premium is more consistent with the role of university-related networks and

signaling than that of human capital.

∗We are grateful to the China Data Center of Tsinghua University and the 90 universities in the survey

for their collaboration. We thank Nick Bloom, Raj Chetty, Aimee Chin, Julie Cullen, Joe Cummins, Gordon

Dahl, Pascaline Dupas, Esther Gehrke, Eric Hanushek, Wei Huang, Craig Mcintosh, Karthik Muralidharan,

Zhuan Pei, Molly Roberts, Noam Yuchtman, and seminar/conference participants at Berkeley, Cornell,

George Washington University, Harvard, NBER, NEUDC, PacDev, University of Houston, Sam Houston

State, Stanford, Tsinghua, UC Riverside, UCSD for their comments. Bingbing Li provides excellent research

assistance.†School of Global Policy and Strategy, University of California, San Diego, and CIFAR, rxjia@ucsd.edu.‡SIEPR, Stanford University, lihongbinsem@gmail.com.

1 Introduction

Access to education, especially elite education, is believed to be one of the most important

channels for elite formation and social mobility in modern societies. For instance, the works

of Pierre Bourdieu emphasize how elite education contributes to the maintenance and repro-

duction of class inequalities, which led to a sociological literature on the social consequences

of elite education institutions (Bourdieu and Passeron 1977; Bourdieu 1988). In economics,

while there exists a burgeoning literature estimating the economic returns to selective educa-

tion (e.g., Dale and Krueger 2002, 2011; Black and Smith 2004; and Hoekstra 2009; Hastings,

Neilson and Zimmerman 2013; Anelli 2016), few studies investigate how the access to elite

education affects elite formation and social mobility.1 In this paper, we not only estimate the

returns to elite education but also attempt to understand the implications of elite education

on elite formation and social mobility using a large-scale dataset we collected ourselves in

China.

Every year, around 10 million students in China take the National College Entrance

Exam – the largest standardized test in the world – in order to get admitted by around 2,300

colleges of different tiers. The exam not only determines whether a young person will attend

a Chinese university, but also which one – attending an elite university is perceived to have

a crucial bearing on career prospects or to provide a ticket to the elite class (Wong 2012).

This perception is not without controversy: due to the importance of family background

for the labor market, some argue that it is illusion rather than reality that a better exam

performance can provide upward mobility and change one’s fate (Bregnbaek 2016).

The exam-based admission of Chinese colleges provides us a Regression Discontinuity

(RD) type laboratory to understand the role of elite education. We refer to elite universi-

ties as those designated by the Chinese government as the first-tier universities across all

provinces in admission, which admit around 5% of the exam-takers in recent years. Elite

universities have a cutoff score, and students just above and below the cutoff score are similar

in most characteristics, but those above are eligible to apply to elite universities. Thus, we

could simply compare outcomes (income and other variables) of students who are just below

and above the cutoff score, which solves the typical empirical issue that students entering

better universities may be different in ability, family background and other characteristics.

We should note that even though the score is the main criterion, it is possible that some

students with scores below the cutoff get accepted with extra points from minor criteria such

as talents in art and sports. It is also possible that some with scores above the cutoff do not

1One exception is a concurrent study on the U.S. (Chetty et al. 2017). We summarize our difference fromtheirs when discussing our contribution to related literature.

1

go to elite universities since the final admission decision is also affected by competition and

students’ preference of majors and location (see Section 2 for discussions). Hence, we have

a fuzzy design with a discontinuity in elite university eligibility around the cutoff scores.

Because there are no existing data we could draw on, for the purpose of our study, we

self-collected systematic data on exam performance and individual outcomes. During 2010-

2015, we designed and conducted annual surveys of college graduates, and collected a total

sample of 40,916 students from 90 universities.2 In our survey, we collected the students’

scores of the college entrance exam, their first job outcomes as well as detailed individual

and parental characteristics. With these data, we are able to examine outcomes of students

with scores close to the cutoffs for elite universities.

We find that elite education in China has a sizable return. Our baseline analysis

focuses on around 10,335 individuals within a bandwidth of 20 points (out of a maximum

of 750 points from four subjects) around the elite university cutoff scores, which is close to

the optimal bandwidth (21) using the method by Imbens-Kalyanaraman (2011). There is

indeed a clear discontinuity around the cutoff: scores above the elite university cutoffs raise

the admission probability by 16 percentage points, about 75% of the mean probability. In

contrast, there is no such pattern with respect to (fake) placebo cutoffs; neither is there

any discontinuity for our balance tests with all individual and family characteristics. Exam

scores above the cutoffs are associated with a wage premium of 150 RMB (USD 25) per

month for a fresh college graduate, which is around 6% of the median monthly wage (2,500

RMB). These together imply an estimate of the monthly wage premium of elite education of

1,000 RMB (USD 160), or about 40% of the median monthly wage. Our results are robust

to alternative bandwidths and specifications of the running function,3 and the potential bias

due to missing wage information seems to be bounded.4

It is worth pointing out that scoring above the cutoff not only changes the probability

of attending an elite university but also other dimensions regarding admission and college

life. In our context, the choices of majors and the relative ranking of students in their

college classes are particularly important. Within a university, different majors also have

different admission scores depending on their popularity. Students often face a tradeoff

between universities and majors, especially for those close to the cutoff. If they choose an

elite university, they are less likely to major in popular fields because, when it is their turn,

2We use “college” and “university” interchangeably throughout the paper.3For instance, Gelman and Imbens (2014) discuss why including higher-order polynomials might bias the

estimates.4To deal with the concern that wages are only observed for those who get job offers, we also present the

bounds on the wage effect following the procedure proposed in Lee (2009) and find this concern is unlikelyto invalidate our approach.

2

those majors may have already been all taken by higher score students. Moreover, those

just above the cutoffs are likely to be the worse students within their college class. While

these dimensions do not affect the internal validity of our empirical strategy (i.e., the wage

premium of elite education is indeed a net outcome of different dimensions), they matter for

the interpretation of our estimate on wage premium, as one may be only interested in the

effect of elite education given everything else equal. When we control for major and relative

class ranking in college, we find a marginally higher elite education wage premium. In other

words, our baseline estimate is likely to be a conservative estimate of elite education premium.

Moreover, we find that the elite education wage premium is unlikely to be explained by

difference in the cost of living.

Market income is only one measure of entering China’s elite class, and there are other

dimensions of being in the elite class that may not be easily monetized. Many of them are

associated with the provision of public goods or perks from government monopolies, such as

getting into the banking industry, finding a job in a state-owned enterprise, and obtaining

hukou (the right to live in a city and enjoy local public goods) of a well-funded municipal. We

also design questions to capture job characteristics beyond wage. We construct a statistical

measure to proxy elite class, namely the ratio between how many people hope to get into

certain occupation (or industry, ownership) and how many actually get into it. The elite

status captured by this ratio is consistent with both the general perception in China about

the elite class and that defined in the sociological literature (see Erikson and Goldthorpe

(1992) for related studies across countries). We find that elite education eligibility does

not increase the probability of entering an elite occupation, elite industry, elite ownership,

living in an elite city, obtaining hukou or any other non-marketized benefit; in contrast,

parental characteristics do. In sum, while the elite education eligibility is associated with a

wage premium, we find no evidence that it necessarily implies one’s entry into the elite club,

defined by job characteristics.

A limitation of our study is that our survey design only allows us to investigate the

wage of the first job for fresh college graduates and it is unclear how important these initial

wages are later in life. To partially address this concern, we supplement our survey data with

job histories of around 300,000 individuals from a major recruitment platform and document

the importance of the first job for future outcomes.

How does elite education eligibility affect social mobility? The question has two di-

mensions. The first is whether elite education eligibility can lift one from a low-income status

to a higher-income status. Not surprisingly, elite education eligibility increases wages and

mobility in this aspect. Moreover, the comparison between the impact of being above the

elite college cutoff score and that of having a rich parent helps us to better understand the

3

degree of mobility created by the exam system in the first dimension. On average, the level-

up effect of elite education is large, moving child income rank up to families with parental

income rank that is two quintiles higher. However, it cannot move an average child in the

bottom quintile (parental) income group to the level of a child from the top quintile income

group. For those students from families of the top 20% income level, even if they are not

eligible for elite education, their probability of being in the top 20% income group is still

higher than those eligible but come from less wealthy families.

The second dimension regards whether elite education eligibility can change the role

family background (such as parental income) plays in determining the job market success of

a person.5 In theory, elite education eligibility can increase (or decrease) intergenerational

mobility if the children with poorer (or richer) parents enjoy higher returns. Empirically,

we find that the intergenerational income (or income rank) link is neutral to elite education

eligibility. In this sense, elite education neither weakens nor strengthens the intergenerational

(relative) mobility; elite education and family background play additive roles in determining

one’s mobility.

Finally, we examine why elite education has a wage premium and consider three ex-

planations: human capital, signaling, and social network. Little empirical evidence exists

to disentangle them due to the challenge of measurement. Exploiting the richness of our

survey data, we proxy them using as many variables as possible. In particular, the existence

of national-level standardized tests in China allows us to have a reasonable proxy for human

capital accumulated in college, and we find no evidence supporting the importance of human

capital in explaining the elite premium. Instead, suggestive evidence appears consistent with

the role of university-related networks and signaling. We further discuss why these channels

can be consistent with our findings on intergenerational mobility.

Our findings contribute to several strands of literature. First, we add to studies es-

timating the economic returns to elite education mentioned above.6 China, as the largest

exam country in the world, provides an interesting case: there are nationally designated

elite universities, and there are cutoff scores for college admissions of all elite universities as

a group. Thus, we could estimate the effect with a large sample of elite colleges, which is

in contrast to most of the existing literature. A few recent studies use the RD method to

estimate the returns to elite education, but almost all these papers estimate the return of

going to one or two specific colleges due to the nature that in most countries each college

5These two dimensions are similar to the definition of absolute mobility and relative mobility in Chettyet al. (2014).

6In the context of China, the 2010 wave of the present data has been used in Li et al. (2012), wherethey control for observables and also find a sizable wage premium of elite education. Our study exploits adifferent empirical strategy that allows us to compare students with similar ability.

4

has its own admission criterion.7

Second, and more importantly, because we designed our own survey and have informa-

tion that is normally unavailable to researchers, we could study how elite education affects

social mobility and shed light on the mechanisms for why elite education matters. Our

findings on intergenerational mobility also add to the growing literature on intergenerational

mobility in economics (e.g., Solon 1992; Chetty et al. 2014; Chen et al. 2015) and sociology

(see Erikson and Goldthorpe 1992, 2002 for related studies). Our approach is complemen-

tary to Chetty et al. (2017), who examine how the parent-child income correlation varies

across colleges in the US. Compared with their study, the advantage of our approach is that

the RD design provides an identification strategy to evaluate elite colleges’ causal effects on

intergenerational mobility. However, the caveat of this approach is that the finding is by

design local to those around the elite university cutoff.8 Different from their results in the

US, we cannot say that elite universities successfully “level the playing field” across students

with different socioeconomic backgrounds. In our case, while elite education does play an

important role in one’s mobility, it does not significantly change the influence of parental

background.

The College Entrance Exam is probably one of the most important institutions in

China, affecting the lives of hundreds of millions of families. A legacy of the imperial exam

institution that was used for over 1,300 years as an elite recruitment channel (Bai and Jia

2016), it gives hope to the Chinese people, especially those not so well-off. Probably due to

the lack of data, there are not many economic accounts of such an important institution.

To the best of our knowledge, ours is the first to systematically study the labor market

consequences of the College Entrance Exam by collecting data on exam scores and exploiting

the cutoffs in recruitment. Our research design and data can be used to study other social

and political implications of elite education.

Section 2 discusses the background and data. Section 3 reports the results on elite

university enrollment, after which Section 4 focuses on wage premium and other aspects of

entry into the elite class. Section 5 centers on social mobility. Section 6 presents evidence

on the mechanisms for the elite wage premium. Section 7 concludes the paper.

7Using a RD strategy, Hoekstra (2009) studies the returns of one elite state university in the US, Anelli(2016) on one elite college in Italy focused on economics/business and Zimmerman (2016) on two elitecolleges in Chile. In particular, Zimmerman (2016) shows that elite peers in two Chilean universities playan important role in explaining the probability of being a business elite. One exception is Hastings, Neilsonand Zimmerman (2013) who investigate the returns by degree programs.

8By design, we are studying students who have achieved the scores around the cutoff. In our data, wedo observe that students from rural areas and poor families are less likely to be able to achieve these scores.This is clearly an important dimension in social mobility omitted by our RD approach.

5

2 Background and Data

2.1 Elite Universities in China

A total of around 2,300 colleges were registered in 2010 in China and the quality of these

colleges varies substantially.9 All of them recruit students based on the score of the National

College Entrance Exam, known as Gaokao in Chinese. They are categorized into four tiers

based on quality and those belonging to a higher tier are afforded priority in admitting

students. A university’s tier can differ across provinces. For instance, a very top university

belongs to the first-tier in all provinces whereas a middle-ranged university are only regarded

as first-tier in certain provinces but not others.

Definition of Elite Universities In this paper, elite universities refer to those falling

under the first tier in the recruitment process across all provinces. Out of all Chinese

universities, 96 universities belong to this category in our study period. Our definition

of elite universities heavily overlaps with two other definitions often used in the literature

(e.g., Carnoy et al. 2013): the so-called Project-211 universities (an abbreviation of the top

100 universities in the 21st century) and the universities under the control of the central

government (111 in total). More specifically, there are 26 elite universities defined by the

national first-tier criterion in our survey: 25 of them belong to the Project-211 group and 21

are under the control of the central government. About 200 universities are locally first-tier

universities (only for certain provinces), which we do not consider as elite since there are no

clear cutoffs and they are of less prestige. The advantage of using the national first-tier to

define elite universities is that they have clear cutoff scores. In recent years, due to college

education expansion in China, over 75% of exam-takers can get college education. However,

only around 5% of exam-takers can get into these elite universities. This ratio also varies

greatly across provinces.10

Tuition, Graduation Rates, and Other University Characteristics It is worth not-

ing that most universities, including all elite universities and second-tier universities (the

focus of this study) are public. Private universities are not very developed in China and the

existing ones are of low quality. Related to their public nature, these universities including

both elite universities and second-tier universities charge similar tuition fees, roughly RMB

9See information from the Ministry of Education on different types of colleges: http://www.moe.edu.cn/publicfiles/business/htmlfiles/moe/s4960/201012/113594.html

10Spatial inequality in access to elite education is an important issue but not the focus of this study, sincewe would like to compare individuals similar to each other in most dimensions including province of origin.

6

http://www.moe.edu.cn/publicfiles/business/htmlfiles/moe/s4960/201012/113594.htmlhttp://www.moe.edu.cn/publicfiles/business/htmlfiles/moe/s4960/201012/113594.html

5,000 a year (USD 750).11 Thus, in our study, we do not need to worry about tuition fees.

In addition, elite universities are of a similar size as non-elite universities in our comparison

group.

Another special institutional feature is that the graduation rates are very high across

all universities in China, with a mean over 95%.12 The college education system in China

is known for being “strict entrance, easy out”. So the probability in graduating varies little

between elite and non-elite universities.

What make elite universities distinct from other universities? They clearly have more

resources and support from the government. For instance, the Project 211 universities ac-

count for only 5% in terms of the number of universities but 70% of all scientific research

funds.13 Naturally, they also attract very different students and teachers. We will present

several university characteristics in our data in Section 2.3.

Going Abroad It is true that some high school students choose to go abroad (especially

to the U.S.) for college. However, this possibility is unlikely to be critical for our study.

First, the decision is usually made before the exam, partly to avoid the exam.14 Even if one

assumes that those who study abroad also take the exam, the share is small. In our study

period, around 0.1 million Chinese students go to the U.S. for college each year, in contrast

with around 10 million taking the exam.

2.2 The College Entrance Exam and Admission

The Admission Process In our study period, the admission process for most provinces

is as follows. All students take the exam in early June in their residence province, which

is written and graded by education authorities in their own province. Then, based on

the distribution of the scores and provincial quotas assigned by the Ministry of Education,

each province announces an admission cutoff score for each layer and each track (social or

natural science).15 With information of the cutoff scores, and also knowing their own scores,

students fill their college preferences for each layer of universities (with universities and

majors). Those above the elite college cutoff scores are eligible to apply for elite universities.

11The lowest-quality ones may charge a higher tuition fee for those who score low in the exam but want acollege degree, but they are not the focus of this study.

12The Beijing-based Mycos Institute estimated in 2011 that China’s college dropout rate is 3%, while theMinistry of Education said that year that the rate is 0.75%.

13See the information from the Ministry of Education: http://www.moe.edu.cn/moe_879/moe_207/moe_235/moe_315/tnull_1469.html.

14See a report on “Applicants say they want to avoid the stress of studying for Chinese college entranceexam”: http://www.chinadaily.com.cn/china/2013-06/07/content_16580626.htm

15The track was determined in the end of the first year of high school and is exogenous to college admission.

7

http://www.moe.edu.cn/moe_879/moe_207/moe_235/moe_315/tnull_1469.htmlhttp://www.moe.edu.cn/moe_879/moe_207/moe_235/moe_315/tnull_1469.htmlhttp://www.chinadaily.com.cn/china/2013-06/07/content_16580626.htm

Finally, universities take turn to recruit based on scores of the applicants. In equilibrium, the

admission bars (lowest admission scores) for different elite universities are different depending

on their popularity among students and competition. First-tier universities recruit first, and

the second-tier universities start only after all first-tier universities finish. Once a student

is admitted, he/she can either choose to go to this university, or decline it and give up

admission to any university for the year.

Across provinces, there are two minor variants of the process: students in Beijing and

Shanghai fill the applications before the exam takes place; students in a few provinces fill

the applications after the exam but before knowing their scores. In all these cases, the cutoff

point is unknown before the exam. Over time, the admission mechanism was transferring

from the Boston Mechanism to one that is similar to the serial dictatorship mechanism in this

period. Again, the exam score is the primary criteria under both mechanisms and the cutoff

determines eligibility. Thus, our empirical strategy is valid despite these complexities.16 As

discussed next, we will compare students around the cutoff within a province-year-(social or

natural science) track.

Features that Matter for Our Research Design The exam system has two features

that are important for our research design.

(i) The scores are only comparable within province-year-track. First, during

2010-2015, 27 out of 31 provinces (except for Jiangsu, Zhejiang, Shanghai and Hainan) use

a scale of 0-750 points based on four subjects,17 while the other four provinces use different

scales. Second, the exam papers are written and graded by each province and a point in

the score in different provinces means differently. Finally, and importantly, the cutoff points

vary greatly across provinces, reflecting spatial inequality in access to elite education. Since

we are interested in comparing individuals similar in all other dimensions except for exam

scores, we will compare students within the same province. We collect the cutoff scores for

each province-year-track from a website specialized for the exam: Gaokao.com. The cutoff

score of elite universities for the 27 provinces using a scale of 0-750 points has a median of

16These variations can have different implications on matching quality. For example, Chen and Kesten(2016) provide an interesting theoretical analysis on the serial dictatorship mechanism. But these variantsare not essential for our research question, as long as students around the cutoff are comparable for a givenmechanism.

17For students in the natural science track, the four subjects are Chinese (with a maximum of 150 points),Mathematics (with a maximum of 150 points), a foreign language (mostly English, with a maximum of150 points) and a combined subject of Physics, Chemistry and Biology (with a maximum of 110, 100, 90respectively). For students in the social science track, the four subjects are Chinese (with a maximum of 150points), Mathematics (with a maximum of 150 points), a foreign language (mostly English, with a maximumof 150 points) and a combined subject of Political Sciences, History and Geography (with a maximum of100, 100, 100 respectively).

8

540.

In our analysis, we will control for province-year-track fixed effects. This takes care of

the four provinces using other exam scales. Our results are also robust to excluding them in

the analysis.

(ii) Our regression-discontinuity is fuzzy. It is possible that some students with

scores below the cutoff get accepted, due to extra scores from other characteristics such as

being an ethnic minority, being a child of a military martyr, or having talents in sports,

music and math etc. These extra scores are orthogonal to the actual scores from the exam.

It is also possible, especially for our focused sample, that some students with scores

above the cutoff do not get into an elite university. To make sure all colleges can recruit

enough students, the cutoff score for each tier (including the elite tier) is set at the level of

105-120% of the admission quota in that tier. Consequently, at least between five to twenty

percent of students passing the elite college cutoff score cannot get into elite colleges. This

means that for those who just pass the cutoff score, the focus of our RD design, their chance

of getting into an elite college is far below 100%. Success for them depends on competition

of a college/major in their preference form. Students could strategize based on historical

data, but it is generally very difficult to predict the level of competition given the large

dispersed potential choices. Each specific combination of college and major normally only

recruits a few students from a province but hundreds of thousands of students are filling in

their preferences at the same time.

Personal preferences about college major and location also matter for the chance of

getting into an elite college. For instance, there is a tradeoff between major and university,

especially for those around the cutoff score. Within a university, different majors also have

different admission scores depending on their popularity. In recent years, economics (includ-

ing finance and business), management and law are popular and hence more competitive in

recruitment. If students with scores just at or slightly above the cutoff choose an elite uni-

versity, they are less likely to major in popular fields because, when it is their time to select,

those majors may have already been all taken by higher score students. Such argument also

applies to preferences for other things such as location of the university.

We will consider these tradeoffs such as major, university location and relative ranking

of students within college class in Section 3.2. We will also check whether personal strategic

choice is a critical concern for our results in Section 3.3.

9

2.3 Chinese College Student Surveys (2010-2015)

The main challenge is to collect data on individual exam scores and link them to labor

market outcomes. We designed and conducted an annual survey of college graduates for this

purpose.

Survey Design and Implementation The data that we use are derived from six waves

(2010-2015) of the Chinese College Student Survey (CCSS), conducted by the China Data

Center of Tsinghua University directed by one of the authors, in collaboration with the

Institute of Higher Education. We randomly selected 100 universities out of all universities

in China by stratifying according to locations (Beijing, Shanghai, Tianjin, Northeastern

China, Eastern China, Central China, and Western China) and tiers of colleges.18 We used

the number of students as weight for each college, meaning that colleges with more students

are more likely to be selected. For the purpose our study, we oversampled elite colleges in

order to get enough students near the elite school cutoff points.

Due to budget and management capacity, the survey was rolled out gradually with

the number of selected colleges listed in Appendix Table A.1. The target was to have all

100 colleges participate in 2013. In practice, 65 colleges participated that year. Due to an

unexpected budgetary cut since 2014, the survey became voluntary, and only those who are

willing and can afford the survey (less than 20 colleges) did it in 2014 and 2015. Across the

six years of survey, 90 colleges out of the 100 participated in at least one of the years.

For each college, we trained a clerk in the student registration office, who helped us

to randomly select a sample of students from the full roaster (population) of the graduating

class. The survey was carried out in May and June each year during 2010-2015. In each of

the participating colleges, we appointed two to three survey administrators, who normally

are in charge of registration, teaching and student affairs. Every year, we trained these

survey administrators from all over the country in Beijing with several days of intensive

meetings. The survey in each college was administered as such. The administrators gathered

all sampled students in a big classroom and let them fill in our paper survey form individually

and anonymously. Students were told at the beginning that these surveys are for studying

issues of the job market for college graduates, and none of their individual information is

disclosed to any party. These filled forms were then coded and mailed to our Beijing office

for data entry and cleaning.

Gathering all sampled students of a college in one location and at the same time

contributes to a relatively high response rate. For the first year (2010), we targeted on

18In the sampling process, we separate these three metropolises (Beijing, Shanghai, and Tianjin) from therest of China because these cities have an extremely large concentration of colleges, especially top universities.

10

selecting 400 students per college and a college on average collected a sample of 319 students,

and our target was cut down to 200 in 2011-2013 due to the rising cost of the survey. For

the last two years in 2014 and 2015, participating colleges agreed to survey 400 students

in each college.19 Across six sounds, the response rates for the elite universities and the

rest are 71.5% and 76.9% respectively and the difference is not statistically significant (with

a p−value of 0.26). These six rounds of survey give us a sample of 40,916 students intheir graduating year. Among them, 34,733 reported complete information on their College

Entrance Exam scores and the provinces where they took the exam.

Appendix Figure A.1(a) plots the spatial distribution of the 90 colleges by province.

As expected, we surveyed more colleges from regions such as Beijing and Shanghai where

there are more colleges. Their students come from all provinces across China, as shown

by Appendix Figure A.1(b), which plots the number of students with exam scores by the

provinces where they took the exams. Note that our strategy is not affected by the selection

of colleges (see discussion below on selection and measurement concerns).

We designed the questionnaire collaboratively with experts in other disciplines (sociol-

ogy and education). One of our aims in the surveys is to gather systematic data on individual

performance at the College Entrance Exam and link it to labor market outcomes. Thus, we

have detailed information on both. We also collected information on family background and

college activities.

Selection and Measurement Concerns We would like to be clear about sample selec-

tion and measurement issues in the survey. Our data on students are not a random sample

of all Chinese college students for two reasons. First, we intentionally over-sampled elite

colleges to capture enough students with elite education. In our sample, 26 out of the 90

universities belong to the elite tier, in contrast to 5% for the population. Second, due to

the nature of voluntary participation for the survey, we do not have 100% compliance from

colleges and students. Note that we do not need a random sample of all college students in

our study. For our analysis, the identification assumption is that those around the cutoff

scores are comparable in all dimensions before entering colleges except for their exam scores.

We will check whether this is the case in the data.

Misreporting and in particular over-reporting exam scores is possible. However, this

concern is unlikely to be critical for our study. Conceptually, we are focusing on exam scores

close to cutoff scores for the elite universities. It is unclear why one would like to misreport

performance just around the cutoff. We expect the concern of misreporting to matter more

19Most of these volunteering colleges intended to collect a sample that is large enough to conduct someanalysis of their own colleges.

11

for the very low and very high scores, which are not our focus. Empirically, we also examine

the density of reported exam scores. Moreover, a clear discontinuity in the first stage (on

how exam scores affect the probability of going to an elite university) around the cutoff

is also reassuring. For instance, Battistin et al. (2009) show that a fuzzy design can still

be applied to identify the causal effect when the first stage discontinuity exists despite the

presence of measurement error in the assignment variable for a subset of observations.

Misreporting of wages is also possible. We exclude 5% of the data (2.5% in both tails

of the wage distribution) in our baseline. The results are robust to keeping all the data

or excluding slightly different percentages of the tail sample. Another selection concern of

wages is that we can observe wages for those who get job offers. To check the importance

of this concern, we will examine the outcomes without job offers and will also estimate the

bounds of the wage effect.

University Characteristics in Our Survey Appendix Table A.2 presents characteristics

of universities in our sample. Consistent with the discussion above, elite universities are

slightly smaller in terms of the number of students but the difference is not significant. Elite

universities clearly have more foreign students and a much larger share of students from

other provinces – consistent with the fact that they attract talents from outside their own

regions.

As expected, there is a big difference in terms of students’ exam scores. The median

score for the elite group is 590 whereas that for the non-elite group is 491. Moreover, the

median first-job wage is about 35% higher for graduates of elite universities.

Data on the College Entrance Exam Scores We collected information on the total

exam score and the score for each subject. Figure 1(a) plots the distribution of the difference

between the reported total scores and the actual cutoffs for elite universities, where the

cutoffs are defined for each province-year-track. The density is weighted by the sampling

weight of universities. Among the 40 thousand graduating students in our surveys, 10,335

of them falls into the 20-point bandwidth centered at the cutoff score. We will use this

20-point bandwidth (roughly 5 points per subject), close to the optimal bandwidth from the

Imbens-Kalyanaraman method (21), as our baseline bandwidth. 21% of the sample within

the 20-point bandwidth went to elite colleges (reported in Table 1 of summary statistics).

For consistency, the summary statistics reported on job outcomes also refer to this sample.

We take a closer look at the sample within the 20-point bandwidth in Figure 1(b). As

shown, there is no significant discontinuity in the reported scores around the cutoff values,

suggesting systematic misreporting around the cutoff is not an important concern. We

12

further plot the distribution of scores for elite universities and non-elite universities separately

in Figures 1(c)-(d). Consistent with the discussion in the background, the probability of

entering an elite university with a score below cutoff is small, while the probability is high

when score is above the cutoff. These figures also suggest that compliance is not 100%, and

thus we have a fuzzy RD design.

Repeated Exam Takers Due to the importance of the exam, some high school students

take it more than once to improve their scores. The share of repeated exam-takers is 30%

in our sample of a 20-point bandwidth. In Section 3, we find that there is no discontinuity

in the probability of being a repeated exam-taker around the cutoff. Moreover, our results

are robust to restricting the sample to the first-time exam-takers.

Data on University, Major and College Academic Performance The survey also

covers information on 13 majors broadly defined. We categorize them into three groups:

STEM (science, engineering, agriculture, medicine and college), Economics (including finance

and business)-Law-Management, and Humanity (philosophy, literature, history, art). As

shown in Table 1, they account for 65%, 24%, and 10% of the students respectively.

The survey also includes questions on the activities and performance of students in

college. We will use these data to shed some light on what elite education brings to students.

For example, performance at the national-level standardized tests in college provides us some

information on human capital acquired in college. In addition, we also ask their perception

about their relative ranking within their college class. We will also examine how it affects

our finding on wage premium.

Data on Job Outcomes Our surveys are conducted in May and June, the last two months

in college for the graduating class. Thus, most of them already have a plan after graduation.

Around 50% of the sample report the best monthly wage offer they get (around 74% of the

sample searched for jobs). The mean and median monthly wage are 2,733 and 2,500 RMB

respectively.

We asked detailed information on job location, industry, occupation, and employer

ownership, which will be used for us to understand what drive the wage premium. Column

(1) of Table A.3 presents the distributions of occupation and industry. As shown, the top

two occupations are professionals (49%) and clerks (28%); the top two industries are manu-

facturing (25%, including construction and mining) and IT broadly defined (15%, including

information, computer and software industry).

A limitation of our survey is that we cannot track an individual once he or she is on

13

the labor market. However, the importance of the first job in lifetime earnings has been

documented by recent studies using administrative data. For instance, Carr and Wiemers

(2016) show a large rank-rank correlation between the first-job wage and that in the future

career. No similar administrative data with job histories are available in China. We collect

job histories of around 300,000 individuals with wages from a major job search platform

(zhaopin.com) and examine the importance of the first job for future outcomes in terms of

wages and elite status of the industry, occupation and ownership.

Data on Parental Characteristics Our survey covers detailed personal and family char-

acteristics including age, gender, residency, parents’ economic and political status etc. We

will check whether students are similar in these characteristics around the cutoffs.

With information on parents, we are also able to link parents’ income and job charac-

teristics (occupation, industry and ownership) with children’s income and job characteristics.

We expect to see intergenerational links in income and job characteristics. Our interest is to

examine whether this intergenerational link gets altered by the exam performance (and elite

education). Related to the literature on intergenerational mobility, we employ both a rank-

rank strategy (examining how the parent income rank affect child wage rank) and a log-log

strategy (estimating the correlation between log parental income and log child income).

3 Exam Scores and Access to Elite Education

We present the impact of exam scores on access to elite education in Section 3.1, with

balance tests of individual and family characteristics. We turn to the possible impacts on

other dimensions such as major in Section 3.2 and discuss whether strategic choice is a

critical concern in Section 3.3.

3.1 The Impact on Access to Elite University

To examine how an individual’s exam score affects elite university enrollment, we use the

following specification:

EliteUnivi,p,y,tr = αEI(Scorei ≥ Cutp,y,tr)+θ1f(Scorei−Cutp,y,tr)+θ2f(Scorei−Cutp,y,tr)×I

+ λp,y,tr + εi,p,y,tr, (1)

where EliteUnivi,p,y,tr is a dummy indicating whether individual i in province p, year y and

(natural or social science) track tr attends an elite university. Scorei indicates individual i’s

14

exam score. The cutoff score for elite universities (Cutp,y,tr) varies by province-year-track.

We include a function f(Scorei−Cutp,y,tr) for the running variable (Scorei−Cutp,y,tr),namely the distance between a student’s score and the cutoff score. To make sure our results

are not driven by certain functional form of f , we employ both the local linear non-parametric

method and the parametric method for this function. In the parametric case, we also allow

the function to differ across the cutoff score by including the linear-interaction or quadratic-

interaction terms between (Scorei − Cutp,y,tr) and I(Scorei ≥ Cutp,y,tr) , and control forprovince-year-track fixed effects (λp,y,tr). All standard errors are clustered at the university

level in the presented results, which are also robust to clustering at the province level or the

score level.

Not surprisingly, there is an obvious discontinuous jump in the probability of admission

to elite universities around the cutoff score. Figure 2(a) plots the probability of attending

an elite university against the running variable by each point of the score in the raw data,

focusing on the range of 20 points below and above the cutoff. The figure shows a notable

discontinuity around the cutoff point: below the cutoff point, the average probability of

attending an elite university is around 0.04 and is fairly stable across scores; above the cutoff

point, the average probability of attending an elite university ranges between 0.19 (for score

at the cutoff) and 0.43 (for 20 points above the cutoff). The difference between one score

above and below the cutoff is sizable (0.15) but still far away from 1. This reflects the

uncertainty around the cutoff discussed in Section 2.1.

Since the cutoff score varies by province-year-track, the effect of one point above the

cutoff in Beijing is not necessarily the same as one point above the cutoff in Shandong. To

allow for such differences, Figure 2(b) controls for province-year-track fixed effects, and the

pattern remains similar.

Regression results confirm the graphical illustrations. In Table 2, we present the

empirical estimates of the impact of being above the cutoff score on entering an elite college

using different methods. Column (1) reports the results from the local linear non-parametric

method. Columns (2)-(5) report the results from the parametric method: columns (3) adds

province-year-track fixed effects; column (4) also controls for a first-order polynomial and

interaction terms; and column (5) adds a second-order polynomial and interaction terms.

As shown, the non-parametric estimate is 0.165 while the parametric estimate is 0.159 after

controlling for polynomial and interaction terms.

Results Using Placebo Cutoffs The sharp discontinuity at the cutoff and the smooth-

ness at other points in Figures 2(a)-(b) already suggest that misreporting around the cutoff

is unlikely to be critical for our findings. Naturally, if we conduct placebo tests using values

15

5-points above and below the cutoffs, we find no similar discontinuity. Moreover, we plot

the patterns using the previous-year and the next-year cutoff scores in Figures 2(c)-(d). The

difference between the actual cutoff and the previous-year cutoff has a mean of -4.5 and a

standard deviation of 30.4. As shown in Figures 2(c)-(d), there is no discontinuity when

replacing the actual cutoff score with either the previous-year cutoff or the next-year cutoff.

These results validate our empirical strategy and also suggest that misreporting near the

cutoff is unlikely to be an important concern.

Balance Tests in Individual and Family Characteristics The underlying assumption

of our strategy is that individuals around the cutoffs are comparable in individual and family

characteristics. To check whether this is the case, we examine whether there is a discrete

jump for a set of individual/family characteristics at the cutoff score.

As visualized in Figure 3, which uses the same specification as in Figure 2(b), there is no

discontinuity for gender, age, being a repeated exam-taker, being in the rural area before col-

lege or father’s/mother’s income. We also conduct the same balance test for other attributes

including whether father/mother having a college degree, and whether father/mother being a

Communist Party member. Results presented in Appendix Table A.4 show no discontinuity

around the cutoff for these characteristics as well.

In addition, we use these observables to predict the wage outcomes and examine

whether the predicted wage exhibits any discontinuity around the cutoff. As shown in

column (10) of Appendix Table A.4, there is no discontinuity around the cutoff either.

3.2 Majors, University Location, and College Class Ranking

As explained in the background, being above cutoff makes a student eligible to apply to

an elite university but he or she still needs to compete with other eligible students in the

admission process. As a result, a student faces tradeoffs such as the choice of university vs.

major, and needs to choose which province to go to college. We examine three important

factors that might also affect wages.

Those slightly above the cutoff are in the worst position for the selection of majors in

the recruitment process of elite universities, while those below the cutoff are in the better

position for non-elite universities. This implies that they are likely to sort into different

majors. For instance, Economics-Management majors are known to be popular in recent

years. Those above the cutoff but are ranked lowest among elite university applicants are

less likely to get into these majors.

Scoring above the cutoff is indeed weakly correlated with college majors, but the cor-

relation is not always significant. On average, the probabilities of majoring in Economics-

16

Management-Law for those above the cutoff and below the cutoff are 21% and 26% respec-

tively. In Appendix Figures A.2 (a)-(c), we plot the difference in the probability of majoring

in Economics-Management-Law, STEM and Humanity. The figures suggest that being above

the cutoff is less likely to major in Economics-Management-Law, which are the most popular

majors in recent years. However, the difference is not always precisely estimated, as shown

in columns (1)-(6) of Table 3.

Scoring above the cutoff also increases the probability of going to a college outside

the home province, as shown by Appendix Figure A.2(d). This pattern is confirmed by

the estimation results in columns (7)-(8) of Table 3: being above the cutoff increases the

probability of attending a university outside the home province by 0.08, around 20% of the

mean (0.38). This finding is consistent with the fact that elite universities are concentrated

in a few cities often located outside one’s home province.

It is conceivable that students who are just above the cutoff are likely to be worse in

terms of academic performance in their class, and this academic ranking might adversely

affect their job outcomes. In the survey, we asked about the ranking within their college

class as perceived by students. There is a systematic upward bias in perception: over 46% of

the students claim to be the top 20% in their class whereas only 5% claim to be the bottom

20% in their class. Even with the bias, there is a systematic difference around the cutoff:

those above the cutoff are less (or more) likely to perceive themselves as the top 20% (or

bottom 20%) in class (shown in Figure A.2 and columns (9)-(10) of Table 3).

We will consider how these factors affect our findings on wage premium in Section 4.

Conceptually, both major and class ranking are likely to bias our estimate of the return to

elite education downward. Thus, the effect of elite education is likely to be larger once we

consider majors and relatively ranking. This is indeed what we find.

3.3 Checking the Importance of Heterogeneous Response

The results in Section 3.2 raise a question on those scoring above the cutoff yet attending

a non-elite university. If the reason for them to attend a non-elite university is due to

competition of their cohort, this is not a concern for our study since their outcome is likely

to be exogenous to their own choice. However, if the reason is due to their own strategic

choice, we need to know whether they are systematically different from those scoring above

the cutoff and attending an elite university. To do so, we examine whether the effect of scoring

above the cutoff on going to an elite university differs greatly with individual characteristics

such as gender, rural residence and family income, and whether individual characteristics

affect the probabilities of majoring in Economics-Management-Law, going to an university

17

out of one’s home province, and being the bottom 20% in college class. Specifically, we

include interactions of I(Scorei ≥ Cutp,y,tr) and individual characteristics in regressions.As shown in Appendix Table A.5, we find no significant heterogeneity conditional on

elite education eligibility (columns (1)-(3)), suggesting that scoring above the cutoff is the

most important criterion for elite education recruitment. The possible tradeoff between elite

education and major, location and college class ranking as discussed in Section 3.2 does not

differ greatly across individual characteristics.

Another concern is that some unobserved characteristics such as risk preference might

affect our results. For example, if, for above-cutoff-score students, those who go to elite

colleges are more (or less) risk averse, then our estimates of the effect of scoring above the

cutoff score (and have elite education) on labor market outcomes could be a result of risk

aversion. We could test whether risk attitude affects our results by using the data in the year

2011, in which we asked two sets of questions to gauge risk attitudes. The first set is to let

students choose between losing (or gaining) 1000 RMB with certainty vs. losing (or gaining)

2000 RMB and 0 with an equal chance. We call a student risk-averse if he or she chooses

certainty for both questions. The second set asks them to evaluate the following statement:

“If you invest, it is more important to ensure a (lower) return with certainty than to take

some risk for the greatest possible return.” We call a student risk-averse if he or she agrees

or strongly agrees with this statement. We define two dummy variables for risk aversion (1

if risk averse) in both the lottery (former setting) and investment cases. The mean values of

the two variables are 16% and 33% respectively.

Regression results presented in Appendix Table A.6 suggest that risk aversion is not

an important factor determining the probability of entering elite universities for students

just above the cutoff score. In column (1), we show that even for data of only one year

(2011), scoring above the cutoff has a similar impact on the probability of entering an elite

university. More importantly, scoring above the cutoff is not systematically correlated with

risk attitudes (columns (2)-(3)). In addition, conditional on scoring above the cutoff, risk

attitudes are not significantly correlated with elite education enrollment either (columns

(4)-(5)). These results should be taken with a grain of salt as we cannot measure the risk

attitudes at the time of application.

The results in Appendix Table A.5 and Table A.6 confirm that being above the cutoff

or not is the primary determinant of elite education enrollment. Conditional on scoring

above the cutoff, there is little heterogeneous response in terms of individual and family

characteristics. Thus, it is reasonable to focus on investigating the average effect of elite

education on labor market outcomes.

18

4 Wage and Elite Formation

We estimate the elite university wage premium in Section 4.1, which speaks to existing

studies on the returns to elite education. Then, to better understand the implications of

elite education, we come to other dimensions of elite formation that are important in the

Chinese context but may not be captured by wage per se in Section 4.2. We further document

the importance of the first job for future outcomes in Section 4.3.

4.1 Estimating the Wage Premium

We first estimate the reduced-form impact of being above the cutoff score of elite universities

on the best wage offer for the first job and evaluate the wage premium of elite education.

Then, we discuss the implications on elite formation of our finding and document the im-

portance of the first-job for future jobs. Before presenting the results on wages, we discuss

whether missing wage information poses a serious selection issue for our estimations.

Who Works? In our data, around 74% of the graduating cohorts have ever searched for

a job. Among those who searched for a job, 74% got at least one offer. As a result, we have

wage information for around half of our sample.

How does being above the cutoff affects the choice to work? We asked students’ post-

graduation plans in the survey. As shown in columns (1) and (2) of Appendix Table A.7,

those above the cutoff are weakly less likely to work after graduation. This difference is not

explained by going to the graduate school (with a mean of 19%) (columns (3)-(4))20 or going

abroad (with a mean of 3%) (columns (5)-(6)). Instead, it is explained by the category of

“unclear plan” (columns (7)-(8)). One explanation is that those above the cutoff may have

more options or higher reservation wages. Relatedly, column (9) shows that those above the

cutoff are weakly less likely to search for a job (3% less likely). Column (10) shows that there

is no significant difference in the probability of getting an offer conditional on job search.

Thus, on the extensive margin, we find that those above the cutoff are slightly less

likely to search for a job. To check whether the selection is critical for our comparison around

the cutoff, we conduct another set of balance tests by restricting the sample to those who

have jobs (and within a 20-point bandwidth). These results are presented in panel (b) of

Appendix Table A.4. Similar to the balance tests in panel (a), the individuals are comparable

20The relationship between enrolling in elite university and going to graduate school is ambiguous. Onthe one hand, those from a non-elite university may be more likely to pursue a graduate degree in an eliteuniversity to improve their job prospects. On the other hand, it is easier for those from an elite universityto continue their graduate studies within the same university. Empirically, we find no discontinuity in goingto graduate schools around the elite university cutoff.

19

in all these attributes. Our results below focus on the sample with wage offers. In addition,

we conduct a bounding exercise to check the quantitative importance of unobserved wages.

The Elite Education Wage Premium The reduced-form specification for wages is as

follows:

(ln)Wagei,p,y,tr = αW I(Scorei ≥ Cutp,y,tr)+θ1f(Scorei−Cutp,y,tr)+θ2f(Scorei−Cutp,y,tr)×I

+ λp,y,tr + εi,p,y,tr, (2)

where the variables are defined in the same way as in equation (1).

There is indeed a discontinuity of log wage at the cutoff score. Figure 4 visualizes

the mean wage and log wage by scores after isolating province-year-track fixed effects. The

pattern on wages is noisier than that for elite university recruitment. Still, we see a notable

discontinuity around the cutoff values.

The graphical results are confirmed by regressions reported in panel (a) of Table 4.

Columns (1) presents the estimate from the local linear non-parametric estimate of the

impact of being above the cutoff score on wage, which shows that having a score above the

cutoffs raises the monthly wage by 122 RMB. When we add province-year-track fixed effects

(column (2)), the first-order polynomial and interaction terms (column (3)), and the second-

order polynomial and interaction terms (column (4)), the effect increases to be 156 RMB

(USD 25), around 6% of the median monthly wage (2,500 RMB). This pattern is confirmed

by the results using log wage as the dependent variable in columns (5)-(8).

Panel (b) presents the first-stage results using the same sample with wage information,

which are comparable with the results in Table 2: being above the cutoff increases the

probability of attending an elite university by around 0.15. The F -statistics are above 27

across specifications, showing that weak instrument is unlikely to be a concern. Algebraically,

the IV estimate is simply the ratio of the reduced-form estimates and the first-stage estimates,

around 1,000 RMB (USD 160) per month, or 40% of the median wage. This finding is

confirmed by the IV estimates reported in panel (c) of Table 4 that range from 33% to 46%.

Of course, one caveat of this IV strategy is the confounding factors of major and ranking

discussed in Section 3.2. We examine how those factors affect our finding on wage premium

and focus on the reduced-form results.

The Impacts of Major, University Location and Relative Ranking To see how our

finding on wage premium is affected by other changes incurred by scoring above the cutoff

point as discussed in Section 3.2, we include 13 major fixed effects, 26 university location-

20

province fixed effects and five within-class ranking fixed effects in the estimation of equation

(2).

Columns (2) and (4) of Table A.8 show that including the major and ranking fixed

effects marginally increases the baseline results. This is consistent with the fact that those

above the cutoff have disadvantages in the choice of major and class ranking. Column (3)

shows that including university location marginally decreases the baseline estiamte, because

elite universities are more likely to be located in more developed provinces. Column (5)

further shows that our baseline findings increase marginally after we control for all sets of

dummy variables.

Finally, column (6) presents the results including university fixed effects. Now, the

effect on wage premium disappears. This finding confirms that the wage premium reflects

an effect of attending different universities.

Results by Bandwidth and by Quartiles We use a bandwidth of 20 points of the score

in our main analysis. A narrower bandwidth implies that individuals are more comparable

but there is more noise in estimation due to a smaller sample size. In Figures 5(a)-(b),

we plot the first-stage and the reduced-form estimates using different bandwidths ranging

from 5 (roughly 1.25 points per subject) to 40 points (10 points per subject), controlling

for province-year-track fixed effects and a quadratic polynomial and interaction terms. As

shown, the results are close to the baseline estimates, suggesting that our findings are not

driven by the IK-optimal bandwidth used in the baseline estimations.

We further check the impacts across the wage distribution. Appendix Figure A.3 plots

the reduced-form results. While the estimates are lower at the tails, the impact is similar

across a large part of the distribution (from the 30th to the 80th percentiles), suggesting

that our finding on the wage premium is not driven by a very limited part of the wage

distribution.

Estimating the Bounds on the Wage Effect As discussed above, we can only observe

wages for those who get job offers and those above the cutoff are slightly less likely to

be observed with wage information. To check how important this selection concern is, we

conduct a bounding exercise following Lee (2009). In our context, the intuition of this

method is to trim the sample below the cutoff (either from above or from below) so that the

share of observations with observed wage information is equal for those above the cutoff and

those below the cutoff. Trimming the sample below the cutoff from below (above) gives us

the lower (upper) bound of the treatment effect of elite education eligibility.

The results are presented in Appendix Table A.9. We focus on the difference in residual

21

log wage after isolating province-year-track fixed effects. Column (1) presents the results

from OLS and columns (2)-(3) present the results with the lower and upper bounds. The

bounded estimate ranges from 0.053 to 0.086, not very far away from our baseline estimate

(0.067). Thus, this selection concern is unlikely to invalidate our approach.

Cost of Living Is it possible that our finding on wage premium is mainly driven by

difference in the cost of living? Our previous analysis controlling for (home) province-year-

track fixed effects has partially dealt with this concern: among those who got job offers, 61%

of the best offers come from employers in their home province. Thus, province-year fixed

effects account for part of the difference in the cost of living across provinces.

We further address this concern by using a provincial-specific CPI available yearly

during 2010-2014 (provided by Brandt and Holz, extending from Brandt and Holz (2006))

to calculate real wages (reported in Columns (1)-(3) of Appendix Table A.10),21 and by

controlling for 31 province-of-job fixed effects (columns (4)-(6)). Regression results suggest

that the wage premium is unlikely to be explained by the cost of living. If anything, the

wage premium appears slightly higher when the potential price effect is controlled for.

4.2 Other Dimensions of Elite Formation

In the context of China, market income is only one measure of entering the elite class.

There are other important dimensions of being in the elite class that may not be easily

monetized, many of which are associated with the provision of public goods or perks from

government monopolies, such as getting into the banking industry, finding a job in a state-

owned enterprise, and obtaining hukou (the right to live in a city and enjoy local public

goods) of a well-funded municipal. We next investigate whether elite education eligibility

affects the entry into elite occupation/industry/ownership and whether it affects non-wage

benefits associated with the job.

For simplicity, we focus on the reduced-form results where the estimates can be in-

terpreted as the impact of elite education eligibility (defined as scoring above the cutoff).

To get a sense of the (IV) impact of elite education, one can divide the estimates by the

first-stage effect (around 0.15). Given a strong first stage shown in Table 4, if the reduced-

form estimate is significant (insignificant), it mechanically follows that the IV estimate is

significant (insignificant).

21The CPI information is available at http://ihome.ust.hk/~socholz/SpatialDeflators.html. To thebest of our knowledge, the information on CPI is only available at the province level.

22

http://ihome.ust.hk/~socholz/SpatialDeflators.html

Elite Occupation, Industry and Ownership We first need to define the elite occupa-

tion, industry and ownership in China. Ex ante, we could define elite occupation, industry

and ownership by the general perception in China. For instance, a job in management is

perceived as an elite occupation; banking is known as an elite industry; and government-

related industries are commonly considered as elite industries. Other than these seemingly

arbitrary definitions, our data in fact allow us to statistically capture these perceptions.

In our survey, we not only asked the actual industry, occupation, and employer own-

ership of the first job offers, but also asked students to report their perceived ideal industry,

occupation, and employer ownership (summarized in column (3) of Table A.3). We call the

ratio of the realized job over the ideal job as the realization-hope ratio, which gives us a

useful measure of the elite status. For instance, if a lot of people hope to work in an industry

but very few can reach it, this industry is likely to be in the elite club.22

The pattern documented from our data is actually consistent with the general percep-

tion in China about elite occupations, industries and ownership, as well as the elite class

defined in the sociological literature. As shown in Appendix Table A.3, occupations like

managerial and administrative positions (in both public or private organizations), business

owners, and military/police, have a realization-hope ratio around 0.1-0.2. In contrast, clerks

and skilled workers have a realization-hope ratio above 4. We define the former group as

elite occupations.23

Similarly, government-related industries have a realization-hope ratio of 0.13; the fi-

nance industry and the education-culture industry have a ratio around 0.4-0.6. In contrast,

mining/manufacturing/construction industries and wholesale/retail have a ratio around 2.

Thus, government-related and finance are reasonable elite industries. As expected, state

and foreign ownerships are elite ownerships. They have a realization-hope ratio of 0.65, in

contrast to a ratio of 2.4 for private ownership.

With these definitions, we examine how elite education eligibility affects the entry into

elite occupation, industry, and ownership. Columns (1)-(3) of Table 5 show that elite educa-

22One caveat of the measure is that the questions on ideal jobs were asked after students have found a job,and thus the answers might be biased by ex post rationalizations. However, this concern does not matterfor our finding unless there is a discontinuity in the bias at the cutoff score, which seems to be a strongassumption.

23In an influential book on class mobility, Erikson and Goldthorpe (1992) categorize social classes intoseven groups (including farmers and manual workers which are not important in our context), where thetop tier refers to “administrators and managers, higher-grade technicians, professionals, supervisors of non-manual workers” and partly coincides with our definite of elite occupation. A nuanced difference is whetherhigher-grade technicians and professionals (whose ratios are 1.4 and 1.3 in our data) should also be includedin the elite class in our context. In our analysis, we focus on the narrow definition of elite occupation (whoseratio is around 0.1-0.2). Scoring above the cutoff weakly decreases the probability of being technicians andprofessionals in our data.

23

tion eligibility has no significant impact on the probability of entering into elite occupations

(managerial and administrative positions, business owners, and military/police), elite indus-

tries (finance, culture-education, governments, and other public sectors), or elite ownership

(state-owned and foreign-owned). The coefficients are of different signs and very close to

zero in terms of elite occupation and elite ownership. Later in Section 5 on intergenerational

mobility, we show that, in stark contrast, parental elite status (defined the same way) does

have a strong influence on the entry into these occupations, industries and ownerships.

Non-Wage Benefits Another measure of elite class is the non-wage benefits on the job.

For instance, the hukou status is an important social indicator for big cities because hukou is

associated with access to public goods such as education, health care, pension etc. Another

example is that some privileged jobs may get more housing-subsidies, which are important

given the rising housing price in China.

Results reported in columns (4)-(7) of Table 5 show that elite education eligibility

does not necessarily bring more job-related benefits. We examine whether the job provides

a local hukou status, whether the job provides housing-subsidy and insurance, and whether

the job is located in major metropolitan cities (Beijing/Tianjin/Shanghai). Once again, the

coefficients are insignificant and of different signs, implying that no systematic impact on

these benefits.

Together with the results on occupation, industry and ownership, these findings suggest

that while elite education eligibility increases wages, it does not promise more benefits in

other dimensions important for the elite class. Section 5 provides more related discussion

when we compare the influence of family background with that of exam performance.

4.3 The Importance of the First Job

In our survey, we only observe the first-job outcomes, but we cannot trace students once

they are on the labor market. Recent studies using administrative data in the U.S. suggests

that the first job plays an important role in determining the future job outcomes (e.g., Carr

and Wiemers 2016, Guvenen et al. 2016). However, we are not aware of similar studies in

China and attempt to provide suggestive evidence here.

To investigate the importance of the first-job in our context, we collect the job histories

of 304,021 individuals from a major online recruitment platform (zhaopin.com). Individuals

there report their monthly wage for each job they have ever had by five categories (1 for below

1000 RMB, 2 for 1000-2000 RMB, 3 for 2001-4000 RMB, 4 for 4001-6000 RMB, and 5 for

6000 RMB and above) and information on industry, occupation and employer’s ownership.

With this information, we can examine the correlation between first job characteristics with

24

future job outcomes. Specifically, we focus on the sample of individuals who have four-year

college education and employ the following specification:

Jobi,u,t = α1Jobi,u,1 + α2Jobi,u,1 ×EliteUniu + λu + γXi,t + γ′Xi,t ×EliteUniu + εi,u,t, (3)

where Jobi,u,t indicates job-related characteristics of individual i who graduated from uni-

versity u in t years after the start of the first job and Jobi,u,1 are the characteristics of the

first job. We also control for university fixed effects (λu), age and gender, and allow their

impacts to vary with elite universities.

The characteristics of the first job are strongly correlated with those in future years.

Columns (1) (and (2)) of Appendix Table A.11 report the correlation between the first

job wage rank and the wage rank 1-5 years (and 6-10 years) after the start of the first

job. Columns (3)-(8) present the correlations in terms of elite industry, occupation, and

ownership. In all cases, the correlations are large and significant. For instance, if one starts

in an elite industry, the probabilities for him or her to stay in an elite industry in the next

1-5 or 6-10 years are 0.61 and 0.47. Therefore, these results suggest that our findings on

the first job also matter for the future. The correlation appears stronger for those from elite

universities in terms of wage rank. But we find no systematic difference between those from

elite universities and the rest in terms of industry, occupation and ownership.

These results suggest that our findings on the first job are also likely to matter for the

future. Of course, these results are not causal since we do not have exogenous variation in

the first job and elite education for this analysis.

5 The Impact on Social Mobility

5.1 Two Dimensions of the Question

How does elite education eligibility affect social mobility? This question has two dimensions.

The first is whether and by how much access to elite education can rasise one’s socioeconomic

status. Our previous findings already suggest the answer: access to elite education increases

wage income and also the rank across income distribution.

The second dimension of this question is less clear: does elite education eligibility

attenuate or strengthen intergenerational mobility? If one believes that elite education in

China levels the playing field and decreases the influence of parental influence, one would

expect to see an attenuation effects. Instead, one may also expect a strengthening effect if

elite education is employed to strengthen the influence of parental income.

25

Appendix Figure A.4 illustrates the two theoretical possibilities, where the x-axis in-

dicates parental income rank and the y-axis indicates expected child’s income rank. In both

cases, being above the elite college cutoff increases the intercept, indicating a positive mobil-

ity in the first dimension. The change in slopes captures the second dimension: being above

the cutoff decreases the intergenerational slope and hence increases the intergenerational

mobility in case (a) and vise versa in case (b).

Below, we examine the pattern in the two dimensions in the data. We measure socioe-

conomic status in terms of income and other important job-related status such as occupation,

industry, and ownership. Following the literature, we use both the rank-rank (e.g., Chetty

et al. 2014) and log-log specifications (e.g., Solon 1992) when examining income mobility.

Again, we focus on the reduced-form estimates, and a significant (insignificant) reduced-form

estimate implies a significant (insignificant) IV estimate, given a strong first stage in Table

4.

5.2 Income Mobility

We first use a rank-rank approach by dividing parental income into five quintiles and rank

it from 1 to 5, with 1 meaning the lowest 20% in the parental income distribution in the

whole sample (with wage information) while 5 the top 20%. The median annual parental

income for the top 20% is around 135,000 RMB, more than 10 folds of that for the bottom

20% group (around 9,000 RMB). Similarly, we divide child wage into five quintiles and rank

it from 1 to 5. We focus on quintiles because we do not have a large sample of individuals

around the cutoff. All the results below are robust to using denciles instead of quintiles.

Moreover, we employ a log-log approach as an additional check.

The correlation between parental income rank and child wage rank, corr(Child Rank,

Parent Rank), captures the intergenerational mobility. Comparing the correlation for stu-

dents above the elite university cutoff and that below can capture the role of elite education

eligibility in altering intergenerational mobility.

Nonparametric Description Before presenting the estimation results, we first present

in Figure 6(a) a nonparametric description for students above and below the cutoff score

separately. For each value (1 to 5) of the parental rank in the x-axis, we plot the mean of

child rank using the same subsample as in the wage premium estimation (i.e., a bandwidth

of 20 points).

Some interesting patterns appear in Figure 6(a). First of all, as expected, there is

a strong and positive correlation between parental rank and child rank for both groups.

Moreover, the correlations are similar, with a slope around 0.2 for both groups. These

26

results suggest that the correlation between parental rank and child rank is not altered by

scoring above the cutoff point of elite universities, or entering an elite university does not

change intergenerational mobility as measured by the generational income link.

The difference lies in the intercept: the group above the cutoff has a higher intercept

(around 0.25 higher). In other words, having a score above the cutoff helps an individual

to move up in the wage ladder by 0.25 quintile and this level-up effect is the same across

parental income rank groups. This level-up effect is large, which is roughly equivalent to the

increase of parental income by about two quintiles. For example, scoring above the cutoff

score can move the child income rank from the first quintile families (the bottom quintile

group of parental income) to the third quintile families (the middle group). However, the

level-up effect is not large enough to move an average child in the bottom group of families

to the top group. For those from the top 20% parental income families, even if they have

scores below the cutoff, their average wage rank is still higher than that of all those above

the cutoff but from families in the other four income groups.

To further shed light on the magnitude, we plot the average probability of getting the

top 20% wage for children by parental income ranks in Figure 6(b). It shows that being

above the cutoffs increases the probability of becoming the top 20% wage earnings among

college graduates across all parental ranks. However, once again, for those from the top 20%

families, even if they score below the cutoff, their probability of earning the top 20% wage

is higher than that of those above the cutoff but from less wealthy families.

Regression Results We can use the following specification to quantify the pattern:

ChildRanki,p,y,tr = β1I(Scorei ≥ Cutp,y,tr)× ParentRank + β2ParentRank + β3I

+ θ1f(Scorei − Cutp,y,tr) + θ2f(Scorei − Cutp,y,tr)× I + λp,y,tr + εi,p,y,tr, (4)

where β3 (together with β1) captures the level effect of being above the cutoff on Child’s

wage rank, β2 (together with β1) measures the intergenerational link of income rank, and β1

captures the difference in the correlation between parental rank and child rank.

There is indeed a strong intergenerational correlation between parental income rank

and child wage rank, as shown by results reported in Table 6. Column (1) shows that the

correlation between parental rank and child rank is around 0.18. Even though the magnitude

is meaningful, one cannot assume that it captures the broad intergenerational mobility in

China: we are studying a selected sample where even those blow the cutoffs attended a

university.24 Column (2) shows that the impact of elite education eligibility is 0.34, which

24As a comparison, using urban household survey in 2004, Chen et al. (2015) examine the correlation

27

is comparable to the effect of an increase in parental income by about 1.8 quintiles. This

impact is sizable but not large enough to lift a child from a bottom 20% family to a top 20%

wage group.

Elite education eligibility does not change the intergenerational correlation, as the

coefficient for the interaction term of above cutoff and parental rank is not statistically

significant in column (3). Column (4) adds province-year-track fixed effects and quadratic

polynomial interactions. As expected, the correlation between parental rank and child rank

becomes smaller. The main result on the interaction effect remains small and insignificant.

These estimation results confirm the nonparametric pattern in Figure 6: the wage premium

of elite education eligibility does not vary much with respect to parental income rank; or put

differently, elite education eligibility does not alter the intergenerational mobility measured

by the correlation between parental income rank and child income rank.

The results from log-log specification imply the same pattern as in the rank-rank

specification. In addition to the rank-rank approach, we also employ a log-log approach

by replacing ChildRank with lnWage, and ParentRank with lnParentIncome. The log-

log estimation results presented in columns (5)-(8) of Table 6 are similar. The correlation

between log parent income and log child wage in a simple regression (column (5)) is around

0.07, and being above the cutoff is associated with a wage premium but does not change

the intergenerational wage correlation (column (7)). The impact of being above the cutoff

is comparable to doubling parental income. When we add province-year-track fixed effects

and quadratic polynomial interactions in column (8), the results are very similar except that

the intergenerational correlation deceases.

5.3 Occupation/Industry/Ownership Mobility

We also examine the intergenerational links in terms of elite occupation, industry, and own-

ership status (defined in the same way as in Section 4.2). Once again, results reported in

Table A.12 illustrate the importance of parental background in determining the job outcomes

of children. As shown, the parental status is significantly correlated with child status in all

three dimensions. The impact of parental background is large compared with mean proba-

bility: having a parent working in an elite occupation, industry, and ownership increases the

probability of entering the elite occupation, industry, and ownership by around 33%, 64%

and 24% respectively (columns (1),(4),(7) of Table A.12). These are larger than the mobility

in terms of income found earlie